Researchers from Zhejiang Gongshang University, the University of Sussex and two other organisations published a preprint on arXiv on 9 November that has not been peer reviewed. It measures how deleting reasoning traces from earlier turns changes how often a reasoning model succeeds.

A reasoning trace is the text a model writes before it acts. In a multi-turn session, the model reads the earlier traces on the next turn. Providers tie a trace to whoever produced it, the authors wrote, so a request routed to another provider arrives without the traces.

One session, three ways

The researchers ran 794 multi-turn tool tasks from the BFCL v4 benchmark on seven open-weight models. They ran each task three times: keeping the traces, deleting the traces from all earlier turns, and deleting traces only from the early part of the session, about 40 per cent of turns. Tool calls with their results and the trace of the current turn were always kept. BFCL’s official checker scored the session’s final state without knowing how the traces had been handled. The researchers repeated the runs with three random seeds. Baichuan-M2-32B and Olmo-3-7B-Think were run on only 200 tasks.

The effect varied by model

The table compares success rates when all earlier traces were kept or deleted.

Source: Jia et al., arXiv 2611.13223. Pooled results across three seeds. A positive difference means deletion lowered the score.

R1-Distill-14B improved from a baseline at which it succeeded on 1.3 per cent of tasks. Olmo’s difference rests on ten and two successes. Model size did not explain the differences: the three 32-billion-parameter models differed by −0.2, +0.7 and +51.3 percentage points.

With Qwen3-32B, the researchers deleted 0–100 per cent of the history on 200 tasks. Success rates at the two extremes differed by 1.5 percentage points. The models sensitive to deletion were not tested this way.

Repeated calls and silent failures

Deletion also changed how the models behaved. After early-session traces were deleted, R1-Distill-14B repeated the same call with the same arguments in 68.7 per cent of sessions, against 18.1 per cent with traces kept. For Olmo, deleting all traces raised the share from 30.2 to 66.5 per cent. Baichuan repeated itself only slightly more, yet still failed. Qwen3-8B made more calls after deletion and succeeded just as often. Whether a model re-derives the deleted reasoning remains open, the authors said.

The authors interpret sensitivity as a property of the model and propose a screen: run the model on 200 tasks with traces kept and deleted. The history is kept if the confidence interval of the difference lies entirely above 10 percentage points. That requires a point estimate above roughly 20 percentage points. The screen would therefore likely class a model whose score drops by, say, 15 percentage points as tolerant of deletion.

The results apply to BFCL-style function-calling tasks. Whether they hold for other tasks is an open question, the authors said. Closed models were not measured. Session logs are available from the authors on request.